About cleaning the training dataset

[AtlantixJJ]

Thanks for sharing of your brilliant work. I plan to implement something similar to your work and I notice that you mention the nico-opendata dataset.

I get the dataset but I found that the dataset is really a mess. Only a few of the images are illustrations of anime characters and most of them are scribbles. I guess you must have cleaned the dataset, would you mind sharing your advice for doing this?